Kala Sundararajan 1, Kelly Lane 2, Wesley Leong 2, Konstantin Shestopaloff 1, Raja Rampersaud 1,3,4, Christian Veillette 1,3,4
1 Arthritis Program, Krembil Research Institute, University Health Network
2 Techna Institute, University Health Network
3 Division of Orthopaedics, Toronto Western Hospital, University Health Network
4 Department of Surgery, University of Toronto
Lumbar spondylolisthesis is a degenerative lower back condition that can cause considerable pain and disability. It is typically treated with a surgical procedure called spine fusion, in which the affected vertebrae are stabilized with rods and screws. The majority of people who undergo this treatment have substantial improvement in pain and function.
However, some patients show minimal improvement, or get worse after surgical treatment. Even experienced surgeons can have difficulty predicting which patients are in this group. Recent research (e.g., the LEAP-OA study at Toronto Western Hospital) has aimed to identify characteristics common to these “non-responders”, with the ultimate goal of predicting a patient’s outcome before surgery so non-responders can be directed to less invasive alternatives. Our objective was to use modern machine learning tools to predict spine fusion outcomes using data collected preoperatively, and to apply the prediction models in a user-friendly clinical decision support tool.
Leveraging data from several prospective cohort studies, we used R software to develop a set of models that use preoperative patient-reported data to predict 1-year postsurgical pain, disability, and patient satisfaction. The models were operationalized as a web service using Microsoft Azure Machine Learning Studio. We also created a dashboard to visualize both patient characteristics and the predictions using Microsoft Power BI. We integrated the web service and Power BI dashboard into the existing UHN-DADOS data collection software.
Results and discussion
The end product was an end-to-end system for clinical data collection, outcome prediction, and visualization. This system is now being piloted at an orthopaedic spine clinic in Michigan. Results of the pilot are pending, but we have shown that it is feasible to develop and implement this kind of clinical decision support tool. We are currently developing a corresponding decision support tool for hip and knee replacement surgery.